Intelligent Advertising in Real Estate: Challenges, Practices, and Insights from Beike

Beike, a leading real‑estate platform, introduced its intelligent advertising system that aims to maximize business goals like MAU, GMV, and conversion rates through a full‑stack DSP platform integrating traffic from major media such as ByteDance and Tencent.

The core advertising workflow includes a seven‑layer funnel: request, bidding decision, winning bid, impression, click, app launch, and final conversion, each requiring pre‑estimation of conversion probabilities.

Key technical components include multi‑task learning models (MMoE) for joint prediction of click‑through and launch rates, a delayed‑feedback model (DFM) for conversion‑rate estimation, and a budget‑smoothing mechanism to control daily spend.

For user targeting, Beike enriches GEO location data by fusing internal building dictionaries with offline DMP and DSP signals, improving accuracy beyond simple IP‑based methods.

To address the long conversion cycle in real‑estate, a feedback‑delay model jointly trains conversion probability and time‑distribution, using exponential decay assumptions and calibrated semi‑conversion signals to reduce label bias.

Budget allocation is driven by ROI maximization, modeling a monotonic but marginally diminishing return curve across channels, and employing global elasticity models with Thompson‑sampling bandits to handle cold‑start and extrapolation challenges.

Finally, Beike evaluates the incremental value of its RTA (Real‑Time Auction) strategies through controlled experiments, comparing various bidding perturbations, media‑direct投, and integrated CVR‑based smart bidding to select the optimal approach.

Overall, the sharing demonstrates how AI‑driven models, multi‑task learning, and ROI‑focused budget allocation can significantly improve advertising effectiveness in the high‑value, long‑cycle real‑estate industry.